基於結構幾何先驗與貝葉斯非局部平均法之SAR影像散斑雜訊高效估計與抑制

Abstract

合成孔徑雷達（Synthetic Aperture Radar, SAR）影像由於雷達訊號的同調特性，固有地存在著乘性散斑雜訊，嚴重降低影像品質並影響後續影像分析與應用。傳統去噪技術雖具計算效率，但往往忽略結構與紋理資訊，導致無法有效地保留影像細節。此外，多數現有的雜訊估測方法依賴特定的參數分布假設或均質區域選取，在面對紋理複雜、結構多變的SAR場景時，這些假設難以成立，因而顯著降低了估測準確性。

本論文提出了一種基於貝葉斯非局部平均（Bayesian Nonlocal Means, BNLM）濾波器與草圖結構先驗（sketch-based geometric priors）的SAR影像散斑雜訊高效估測與抑制方法。在雜訊估測方面，研究結合了離散小波轉換（Discrete Wavelet Transform, DWT）、適應性局部變異數估測與多項式迴歸，成功地將雜訊估測的平均誤差大幅降低至2.66%，顯著優於傳統的Gaussian-Hermite方法與Generalized Gamma Distribution方法。

在雜訊抑制方面，我們所提出的改良型BNLM框架包含兩項關鍵改進：(1)以Wiener filter為基礎之先驗估測方法，能有效降低估測偏差並更精確地反映局部變異數 (2)透過草圖結構導引之異向性高斯核函數，能更精準地保留影像中的邊緣與紋理結構，本研究所提出之方法在量化指標PSNR、SSIM、運算效率上均有明顯的提升。

總體來說，本研究所提出之方法，在公開的SAR機場資料集上進行的實驗驗證顯示，在SAR影像散斑雜訊估測與去噪中的準確性、視覺真實性及運算效率上皆有顯著提升，未來更可廣泛應用於衛星影像即時處理等多元化遙測領域。

Synthetic Aperture Radar (SAR) imagery inherently suffers from multiplicative speckle noise, which significantly deteriorates image quality and complicates subsequent analysis tasks. Conventional denoising approaches typically overlook critical structural and textural details, while prevalent noise estimation techniques are constrained by oversimplified parametric assumptions, leading to suboptimal accuracy in complex SAR scenes.

This study presents a robust framework for SAR speckle noise estimation and suppression, integrating Bayesian Nonlocal Means (BNLM) filtering with sketch-based geometric priors. In the noise estimation phase, a combination of Discrete Wavelet Transform (DWT), adaptive local variance estimation, and polynomial regression markedly reduces the estimation error to approximately 2.66%, surpassing conventional methodologies.

For noise suppression, the proposed enhanced BNLM method incorporates a Wiener filter-based prior to accurately model local variance and employs an anisotropic Gaussian kernel driven by geometric structural cues. Empirical evaluations utilizing public SAR datasets demonstrate substantial enhancements in Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and computational efficiency, underscoring the method's applicability for real-time satellite imagery processing.